A national assessment of the sensitivity of Australian runoff to climate change

Abstract

The relationship between catchment rainfall, evapotranspiration and runoff can be exploited to assess climate risk to water resources. National data regarding climatology and runoff were used to estimate the sensitivity of regional runoff to projected changes in precipitation and evaporation. These sensitivity factors were integrated with patterns of climate change from 12 different global climate model (GCM) simulations to project future annual runoff sensitivity per degree of global mean temperature change. Divergent runoff sensitivities were identified depending upon the selected GCM. Averaging among GCMs resulted in a robust pattern of runoff sensitivity suitable for estimating future climate risk. Copyright © 2008 Royal Meteorological Society

1. Introduction

The security of Australia's water resources has been identified as one of the most vulnerable aspects of the Australian environment to the effects of climate change (Allen Consulting, 2005). This vulnerability emerges from a combination of factors including the aridity of the continent, high inter-annual and inter-decadal variability in climatic conditions, the intensive use of existing resources, potentially large changes in future climatic conditions, and growing human demand (Jones et al., 2008). Given this known vulnerability, there is acute interest and value in better characterising the response of the continent's hydrology to the effects of climate change.

A number of hydrological models, such as the integrated quality and quantity model (IQQM), SIMHYD, and the Australian water balance model (AWBM), have been applied in the assessment of the future implications of climate change for Australian runoff and streamflow in select catchments. Owing to model complexity and short time-steps of simulations, the parameterisation of such models is somewhat data intensive. Hence, past work has focussed on high-value catchments, such as those within the Murray-Darling Basin (Jones and Page, 2001; Beare and Heaney, 2002; Bates et al., 2003), or has attempted to sample from a small number of catchments across the continent (Chiew and McMahon, 2002; Chiew et al., 2003; Jones et al., 2006) under a limited suite of assumptions about the future climate. As a consequence, a comprehensive picture of potential climate risk to water resources across the Australian landscape that accounts for both climatological and hydrological uncertainties is yet to emerge.

Recently, a simple but robust method for estimating changes in annual runoff at the national scale has been developed (Jones et al., 2005, 2005; Jones and Durack, 2005; Jones and Preston, 2008). The approach is based upon the quantification of the statistical relationship between rainfall and evapotranspiration and the subsequent response of catchment runoff as simulated by multiple hydrological models operating at the catchment scale (Jones et al., 2006). Application of these simple relationships enables one to estimate a catchment's runoff sensitivity to changes in rainfall and evapotranspiration and, subsequently, to estimate changes in future runoff in response to scenarios of climatic change (refer also Wigley and Jones, 1985; Arnell, 1992).

Here, this approach is expanded to estimate runoff sensitivity for the entire Australian continent. Gridded data sets representing the sensitivity of runoff across the landscape to changes in rainfall and evapotranspiration are developed from data generated from prior national-scale climatological and hydrological assessments. These sensitivities are then integrated with Australian patterns of climate change from a range of different global climate model (GCM) simulations to estimate model-specific projections of runoff sensitivity in response to climate change.

2. Methods

The impacts of climate change on Australian catchment basins were quantified through the use of a simple hydrological rainfall-runoff model. Runoff is one of the most sensitive and robust measures of climate change's influence on surface-water resources, as it integrates the effects of rainfall and evapotranspiration as well as landscape characteristics but lies upstream of hydrological management and human use, which affects stream flows and storage levels.

The response of runoff in Australian catchments to climate change has been previously estimated as the change in mean annual runoff produced by a specific hydrological model in response to changes in mean annual precipitation and potential evapotranspiration (E_p). As such, annual runoff can be expressed as

• (1)

where δQ is the change in mean annual runoff, δP is the change in mean annual precipitation, δE_p is the change in mean annual potential evapotranspiration (all measured in percent), and α and β are constants. α is a measure of the sensitivity of catchment runoff to changes in P and β to changes in E_p. This simple relationship has been parameterised based upon simulations with the IQQM of the effects of climate change scenarios on the Macquarie and Fitzroy Rivers in Australia (Jones and Page, 2001; Jones et al., 2005,2005). IQQM is a water balance model that can represent reservoirs, wetlands, and surface-water/groundwater interactions for both natural and regulated systems. For the Macquarie River, the simple hydrological model was capable of reproducing IQQM estimates of changes in water storages (r² = 0.98), wetland inflows (r² = 0.98), and irrigation allocations (r² = 0.98) in response to different climate change scenarios with standard errors of ± 2.0%. In the Fitzroy River, the simple model was able to reproduce mean annual flow (r² = 0.97) with a standard error of ± 2.0%.

Applying this relationship on a national scale necessitated the estimation of the continental-scale spatial distribution of α and β. Jones et al. (2006) found that the average sensitivity of catchment runoff to rainfall and evaporation is correlated with the catchment runoff coefficient (Q_c; the ratio of annual runoff to rainfall as derived from monthly data; 1980–1999). As a consequence, given the data for Q_c (see below), it is possible to calculate both α and β. Values for α and β for 22 catchments across Australia were calculated from two different hydrological models (SIMHYD and AWBM) by Jones et al. (2006), with values for α and β ranging from 1.83 to 4.06 and − 1.15 to − 0.07, respectively. Least-squares linear regression was subsequently used to calculate generalisable statistical relationships between Q_c (expressed as a percentage of annual rainfall) and α and β (Jones et al., 2005,2005; Preston and Jones, 2008):

• (2)

and

• (3)

As apparent from Equations (2) and (3), variation in α has the greatest bearing on δQ. Although the r² statistic for β is statistically marginally insignificant, in the calibration of the simple model against IQQM simulations for the Macquarie and Fitzroy Rivers, the incorporation of evaporation was necessary to maximise the goodness-of-fit. This is due to the inclusion of evaporation as a fundamental climate variable within IQQM and other process-based hydrological models (Jones and Page, 2001; Raupach et al., 2001a,b; Jones et al., 2005,2005). Nevertheless, on the basis of Equations (2) and (3), values for β are relatively small compared to α. Hence, δP is inherently the driving variable in the simple hydrological model (δQ is up to 3.5 times more sensitive to δP than δE_p), making the model relatively insensitive to changes in E_p.

To generate continental-scale estimates of Q_c, monthly runoff and rainfall data were obtained as 5-km gridded data sets from the Australian Natural Resources Data Library (NRDL, 2006a,b). The rainfall data were originally derived from observations made by the Australian Bureau of Meteorology (BOM), interpolated to daily time-step and a spatial grid of 0.05° by the Queensland Department of Natural Resources (QDNR) (see also Jeffrey et al., 2001). Runoff data were originally derived for the period 1980–1999 using the BiosEquil model (Raupach et al., 2001a,b), and represent the sum of rainfall plus irrigation from storages minus evapotranspiration. Continental-scale estimates of Q_c at 5-km resolution were generated by taking the ratio of runoff to rainfall using ArcGIS Desktop 9.2. These data were subsequently applied in Equations (2) and (3) to derive 5-km gridded datasets for α and β across the Australian landscape (Figure 1). These data indicate that the sensitivity of runoff to rainfall (α) is highest in the arid interior and lowest in the high-rainfall areas of the far north, southeast, and southwest. Similarly, sensitivity to evaporation (β) exhibits a similar pattern, although the sign of the runoff response to a marginal increase in evaporation is opposite to that of rainfall.

Figure 1. Spatial patterns of α and β for the simple runoff sensitivity model. α and β were based upon monthly mean runoff expressed as a percentage of monthly rainfall. (^†Correction made here after initial online publication.)

Scenarios for future percentage changes in annual, DJF, MAM, JJA, and SON rainfall and evaporation based upon the aforementioned 12 GCM simulations were obtained from the OZCLIM scenario generator as 25-km grids representing patterns of change per 1 °C of global warming relative to 1990 (Table I; Page and Jones, 2001). OZCLIM data were originally derived by applying a pattern scaling technique to GCM simulations obtained from the Program for Climate Model Diagnosis and Intercomparison (Whetton et al., 2005), and patterns were regridded to 25 km using a simple interpolation technique. Twelve different patterns were used based upon a range of GCM simulations, representing a diverse array of regional representations of the effects of global greenhouse gas emissions on Australian climate.

Table 1. Climate model simulations analysed in this report. Further information about the non-CSIRO simulations may be found at the IPCC Data Distribution Centre (see also Whetton et al., 2005). SRES refers to the IPCC's Special Report on Emissions Scenarios (IPCC, 2000)

Centre	Model	Emissions scenarios post-1990 (historical forcing prior to 1990)	Years	Horizontal resolution (km)
DARLAM and CC are regional climate models.
Canadian CC	CGCM1	IS92a	1900–2100	∼400
Canadian CC	CGCM2	IS92a	1961–2100	∼400
Canadian CC	CGCM2	CO2 + aerosol SRES, A2	1900–2100	∼400
Canadian CC	CGCM2	CO2 + aerosol SRES, B2	1900–2100	∼400
CSIRO, Australia	Mark2	IS92a	1881–2100	∼400
CSIRO, Australia	DARLAM	IS92a	1961–2100	125
CSIRO, Australia	Mark3.0	SRES A2	1961–2100	∼200
CSIRO, Australia	CC50	SRES A2	1961–2100	50
DKRZ, Germany	ECHAM4/OPYC3	IS92a	1990–2100	∼300
Hadley Centre, UK	HadCM3	IS92a	1861–2099	∼400
Hadley Centre, UK	HadCM3	CO2 + O3 + aerosol, SRES, A2	1950–2099	∼400
Hadley Centre, UK	HadCM3	CO2 + O3 + aerosol, SRES, B2	1950–2099	∼400

As noted in Jones et al. (2005,2005), both rainfall and runoff have strong seasonal signals in Australia. For example, peak rainfall in the northern half of the continent occurs during the warm season monsoon, from December to March, while peak rainfall in the southern half falls between late fall and early spring (although there is significant geographic variability). Failure to account for this can potentially contribute to a relatively high rate of errors in runoff estimates (Jones et al., 2005,2005). For example, in the absence of seasonal weighting of annual runoff estimates the simple model performed relatively poorly in reproducing mean annual flows of the Fitzroy River (r² = 0.88; standard error = 4%; Jones et al., 2005,2005). A simple correction whereby average seasonal quarterly rainfall and evaporation were weighted by average quarterly runoff lagged by one month (i.e. DJF rainfall was weighted according to the proportion of annual runoff occurring in JFM) was found to improve the goodness-of-fit of the simple model to IQQM (r² = 0.97; standard error = 2%). This same seasonal weighting scheme was applied in the current study (Figure 2), and a sensitivity analysis of seasonal weights is reported elsewhere (Preston and Jones, 2008). Weights were calculated based upon the 5-km monthly gridded runoff data obtained from the NRDL. These weights were then applied to climate scenarios of quarterly changes in precipitation and evapotranspiration and quarters were summed to obtain a weighted annual change. The runoff-weighted annual changes in rainfall and evapotranspiration from each GCM simulation were applied to α and β in Equation (1) to calculate runoff per degree of global mean temperature change.

Figure 2. Geographic distribution of seasonal weights used to generate runoff-weighted annual changes in rainfall and evaporation. Weights were based upon the ratio of seasonal quarters of runoff to the annual average. A one-month lag was used in the application of weights, such that the DJF weights are based upon JFM runoff

3. Results

Inspection of runoff sensitivities from the various GCM simulations reveals highly divergent estimates of the regional implications of climate change for Australian runoff and water resources (Figure 3). For example, CC50, ECHAM4 and MK3 suggest significantly reduced rainfall in response to increases in global temperature, particularly in Western Australia. In contrast, other models, such as DARLAM, project wetter futures, with runoff increasing. The various climate models also indicate that runoff sensitivities, and therefore future changes in runoff, can potentially be quite large, with increases or decreases of ± 30% for extensive regions with a 1 °C increase in global mean temperature, and even larger changes for more isolated areas.

Figure 3. Geographic distribution of runoff sensitivity for the 12 global and regional climate models applied in the current study. Patterns represent the percentage change in annual runoff in response to a 1 °C increase in global mean temperature

Owing to disparities in results from various GCM simulations, patterns of runoff change from individual GCMs were averaged using ArcGIS Desktop 9.2, and the standard deviation was calculated to assess runoff uncertainty. This ensemble pattern indicated three regions of high runoff sensitivity in response to climate change. First, coastal West Australia, from Pilbara to the State's southwest, was associated with significant reductions in runoff on the order of 10–20% per 1 °C increase in global mean temperature. Second, Southeast Queensland on the continent's east coast also was associated with similarly sized reductions in runoff. Third, the arid interior of southern Australia, from central West Australia east to southern Northern Territory and South Australia was associated with moderate increases in runoff of less than 10% per 1 °C increase in global mean temperature. Similar increases were projected for Tasmania. Other areas of the continent indicated small changes within ± 5%.

Overall, standard deviations for mean runoff sensitivity were relatively large throughout much of the continent relative to the projected runoff changes (Figure 4). Regions associated with low standard deviations for runoff sensitivity were confined to the continent's far north, southeast, and deep southwest. Meanwhile, standard deviations were highest in northwest West Australia around the Pilbara region, as well as southeast West Australia and isolated areas in the Northern Territory, South Australia, and eastern Tasmania (Figure 4).

Figure 4. Mean and standard deviation of runoff sensitivities from the 12 global and regional climate models applied in the current study

4. Discussion and conclusions

The simple hydrological model presented here represents a useful approach for characterising the first-order sensitivity of terrestrial runoff to annual climate variability and change. The continental-scale hydrological sensitivity of Australia indicates the potential for significant adverse impacts to regional water resource security in a changing climate (Jones et al., 2008; Preston and Jones, 2008). For example, given 1 °C of global warming, average model results suggest the potential for annual reductions in runoff on the order of 20% in coastal West Australia and Southeast Queensland, although the seasonal distribution of runoff changes varies significantly between these two locations (see Preston and Jones, 2008). Given the estimates of increases in global mean temperature of up to 6.4 °C by 2100 (IPCC, 2007), the runoff sensitivities indicated by the model suggest a high degree of climate risk and substantial water management challenges, should the more pessimistic scenarios of future global temperature increases and drying projected by the balance of models for the continent's west, east, and southeast be realised.

Projections of runoff changes merit cautious interpretation because of the model's simplicity as well as the disparities in rainfall, and subsequently runoff, projected by different GCMs. For example, the spatial patterns of future climate and runoff generated by the CC50 and DARLAM models are in direct opposition (Figure 3), with the former suggesting significant drying across central to southern Australia while the latter suggesting increasing rainfall and runoff. While averaging across different GCMs is one tool for addressing such uncertainty, this is dependent upon the arguable assumption that the different GCM representations of future climate are equally likely (CSIRO and BOM, 2007).

The sensitivity of runoff at any given locale or region to changes in rainfall and evapotranspiration is also the net result of a complex set of characteristics including soil moisture, topography and geology, and vegetation. Such uncertainty in hydrological sensitivity is not comprehensively explored here. However, previous investigators have noted that such uncertainties are small relative to those associated with future climate conditions (Arnell and Liu, 2001; Jones et al., 2006). In addition, the hydrological model relies upon linear relationships between changes in global mean temperature, regional climate changes, and catchment runoff. Chiew (2006) has noted the potential for linear pattern scaling methods to overestimate runoff reductions in catchments where climate models project increases in extreme rainfall events (which are associated with a high magnitude of runoff) despite reductions in average annual rainfall. Meanwhile, the dynamics of hydrological systems are known to display non-linear responses and thresholds (Arnell and Liu, 2001), particularly in response to large perturbations from the mean state. The potential shortcomings of linear scaling become apparent when one considers the implications of the estimated runoff sensitivities reported here in light of high magnitudes of future increases in global temperature. Given a 5 °C increase in the global mean temperature, the runoff sensitivities reported here suggest annual runoff in a number of locations in Australia effectively stops. Such implausible results indicate linear scaling may result in overestimation of runoff changes.

Despite these limitations, the simplicity of the hydrological model presented here enables runoff sensitivity to be estimated over large geographic areas at relatively high resolution, even in the absence of detailed hydrological information. This facilitates its use in a number of assessment applications. The model can be used for rapid scoping of the risk of climate change across the Australian landscape as well as individual catchments. Meanwhile, it enables comparison of runoff sensitivities associated with different GCMs and testing the implications of different configurations of GCM ensembles on climate change impacts (Preston and Jones, 2008). It also provides a tool for conducting stochastic uncertainty analysis with modest computing resources to explore the uncertainty space of future climate change on Australian runoff.